Title: A comparison of matching and machine learning-based covariate adjustment
Authors: Dylan Small - University of Pennsylvania (United States)
Luke Keele - University of Pennsylvania (United States) [presenting]
Abstract: Matching algorithms have become one frequently used method for statistical adjustment under a selection on observables identification strategy. Matching methods typically focus on modeling the treatment assignment process rather than the outcome. Many of the recent advances in matching allow for various forms of covariate prioritization. This allows analysts to emphasize the adjustment of some covariates over others, typically based on subject matter expertise. While flexible machine learning methods have a long history of being used for statistical prediction, they have generally seen little use in causal modeling. However, recent work has developed flexible machine learning methods based on outcome models for the estimation of causal effects. These methods are designed to use little analyst input. All covariate prioritization is done by the learner. In this study, we replicate five published studies that used customized matching methods for covariate prioritization. In each of these studies, subsets of covariates were given priority in the match based on substantive expertise. We replicate these studies using BART, a machine learning method that has been used for causal modeling. We record differences in both point estimates, confidence interval length, and sample trimming.